Leak Suggests GeForce RTX 4070 May Challenge The Mighty $2,000 RTX 3090 Ti
While CPU performance typically grows in terms of percentage points with each successive generation, GPU performance increases tend to come in the form of whole-number multipliers. There's a few reasons for this, but the biggest one is simply that GPU workloads are a massively-parallel task, and it's easy to accelerate them by simply piling on more functional units in your parallel processor.
So it goes that folks are expecting massive performance gains with the next GPU generation, but this next one might really be a doozy. We're already in the midst of a GPU generation with immense distance between the bottom-end and top-end GPUs from the same family, and it looks like the next-generation cards will eclipse this generation's models in both performance and power consumption.
Meanwhile, AD104 is the third-largest Ada Lovelace processor, behind AD103 and AD102. Going by the leaked information we've seen previously, it will be the GPU that powers the GeForce RTX 4070 and its associated products. WIth that in mind, kopite7kimi is essentially implying that the GeForce RTX 4070 (or perhaps an RTX 4070 Ti or such) will be as fast as a card based on GA102.
The simplest assumption would be to assume that kopite7kimi is talking about the fully-enabled versions of both GPUs. If we look at their specifications, that seems a little unlikely, but not impossible. Ada Lovelace is supposed to clock much higher than Ampere thanks in part to a significant die shrink. A much higher core clock and the massively-increased cache capacity of Ada Lovelace could make up the difference in functional units. If that's the case, it would be impressive given that the RTX 3090 Ti is a $1,999 GPU (MSRP), whereas the GeForce RTX 4070 will probably land somewhere closer to $499.
We really don't know what specific graphics cards kopite7kimi is comparing, though. As noted above, these GPUs get used (in further and further cut-down forms) for a variety of graphics card SKUs. It's possible that the fully-enabled AD104 only matches up to the original GeForce RTX 3080 10GB card. That's still based on GA102, so we're back to my earlier statement about the breadth of performance in this single generation of GPUs.
Unfortunately, given that NVIDIA and its partners seem loathe to launch a new model of GPU any time soon, we'll probably have at least four or five months to wait before we can lay hands on Lady Lovelace and discern the real story. Of course, we'll let our wonderful readers know as soon as we do.